search

xiaoachen98 / Open-LLaVA-NeXT

An open-source implementation for training LLaVA-NeXT.
240 stars 10 forks source link

Inverse Loss Spike issue #16

Closed LaBaZh closed 1 month ago

LaBaZh commented 1 month ago

When finetuning pretrained Open-LLaVA-Next on the mixture data, I encoutered inverse loss spike issue. Is such issue caused by the mixture form of data? Is it ok to have such curve during the finetuning?

image
Cooperx521 commented 1 month ago

It is very normal and also occurs in the original llava1.5, likely due to the pure text content, although I have not validated it.